Update app.py

app.py

@@ -8,13 +8,10 @@ API_TOKEN = os.getenv("API_TOKEN")
 
 def get_ai_response(message, history):
     """Fetch AI response from the API using the modern messages format."""
-    messages = [
-        {"role": "system", "content": "You are a helpful assistant."}
-    ]
+    messages = [{"role": "system", "content": "You are a helpful assistant."}]
     # Build the API history using all prior complete pairs
     for user_msg, ai_msg in history:
-        # Only include completed exchanges (skip "pending")
-        if ai_msg != "pending":
+        if ai_msg != "⏳ Thinking...":
             clean_ai_msg = re.sub(r'<details>.*?</details>', '', ai_msg, flags=re.DOTALL)
             clean_ai_msg = re.sub(r'<[^>]*>', '', clean_ai_msg)
             messages.append({"role": "user", "content": user_msg})
@@ -54,22 +51,32 @@ def convert_reasoning_to_collapsible(text):
     return html_response
 
 def add_user_message(message, history):
-    """Immediately add the user's message with a 'pending' assistant reply."""
+    """Immediately add the user's message with a '⏳ Thinking...' assistant reply."""
     if history is None:
         history = []
-    history.append((message, "pending"))
-    return history
+    history.append((message, "⏳ Thinking..."))
+    # Return both updated state and chatbot messages
+    return history, history
 
 def generate_response_from_history(history):
     """Generate the assistant's reply and update the last pending message."""
     if not history:
-        return history
-    # Get the last user message from history (which is paired with "pending")
+        return history, history
+    # Get the last user message (which is paired with "⏳ Thinking...")
     last_user_message = history[-1][0]
-    # Get the full conversation (ignoring the pending reply)
-    ai_response = get_ai_response(last_user_message, history)
+    # Build API history excluding pending messages
+    api_history = []
+    for user_msg, ai_msg in history:
+        if ai_msg != "⏳ Thinking...":
+            clean_ai_msg = re.sub(r'<details>.*?</details>', '', ai_msg, flags=re.DOTALL)
+            clean_ai_msg = re.sub(r'<[^>]*>', '', clean_ai_msg)
+            api_history.append({"role": "user", "content": user_msg})
+            api_history.append({"role": "assistant", "content": clean_ai_msg})
+    # Append the last user message to fetch the assistant's reply
+    api_history.append({"role": "user", "content": last_user_message})
+    ai_response = get_ai_response(last_user_message, api_history)
     history[-1] = (last_user_message, ai_response)
-    return history
+    return history, history
 
 # Modern CSS for a clean UI
 custom_css = """
@@ -93,7 +100,8 @@ with gr.Blocks(css=custom_css, title="Reka Flash 3") as demo:
             submit_btn = gr.Button("Send", size="lg")
         clear_chat_btn = gr.Button("Clear Chat")
     
-    chat_state = gr.State([])    # Current chat history
+    # State management for chat history
+    chat_state = gr.State([])
 
     js = """
     function() {
@@ -118,7 +126,7 @@ with gr.Blocks(css=custom_css, title="Reka Flash 3") as demo:
     }
     """
 
-    # First, add the user message immediately (with pending), then update it with the generated response.
+    # First, add the user message with a pending reply, then update it with the actual response.
     submit_btn.click(
         add_user_message,
         [message, chat_state],
@@ -128,12 +136,12 @@ with gr.Blocks(css=custom_css, title="Reka Flash 3") as demo:
         chat_state,
         [chat_state, chatbot]
     ).then(
-        lambda: "",  # Clear the input box
+        lambda: "",  # Clear the input box after processing
         None,
         message
     )
 
-    # Allow pressing Enter to submit
+    # Enable pressing Enter to submit
     message.submit(
         add_user_message,
         [message, chat_state],
@@ -154,6 +162,7 @@ with gr.Blocks(css=custom_css, title="Reka Flash 3") as demo:
         [chat_state, chatbot]
     )
 
+    # Load JavaScript to enable HTML rendering in chatbot messages
     demo.load(
         fn=lambda: None,
         inputs=None,